xinference 1.0.1__py3-none-any.whl → 1.2.1__py3-none-any.whl

xinference/model/audio/fish_speech.py

@@ -11,10 +11,8 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-import gc
 import logging
 import os.path
-import queue
 import sys
 from io import BytesIO
 from typing import TYPE_CHECKING, Optional
@@ -60,6 +58,7 @@ class FishSpeechModel:
         self._device = device
         self._llama_queue = None
         self._model = None
+        self._engine = None
         self._kwargs = kwargs
 
     @property
@@ -72,6 +71,7 @@ class FishSpeechModel:
             0, os.path.join(os.path.dirname(__file__), "../../thirdparty/fish_speech")
         )
 
+        from tools.inference_engine import TTSInferenceEngine
         from tools.llama.generate import launch_thread_safe_queue
         from tools.vqgan.inference import load_model as load_decoder_model
 
@@ -81,6 +81,11 @@ class FishSpeechModel:
             if not is_device_available(self._device):
                 raise ValueError(f"Device {self._device} is not available!")
 
+        # https://github.com/pytorch/pytorch/issues/129207
+        if self._device == "mps":
+            logger.warning("The Conv1d has bugs on MPS backend, fallback to CPU.")
+            self._device = "cpu"
+
         enable_compile = self._kwargs.get("compile", False)
         precision = self._kwargs.get("precision", torch.bfloat16)
         logger.info("Loading Llama model, compile=%s...", enable_compile)
@@ -102,102 +107,10 @@ class FishSpeechModel:
             device=self._device,
         )
 
-    @torch.inference_mode()
-    def _inference(
-        self,
-        text,
-        enable_reference_audio,
-        reference_audio,
-        reference_text,
-        max_new_tokens,
-        chunk_length,
-        top_p,
-        repetition_penalty,
-        temperature,
-        seed="0",
-        streaming=False,
-    ):
-        from fish_speech.utils import autocast_exclude_mps, set_seed
-        from tools.api import decode_vq_tokens, encode_reference
-        from tools.llama.generate import (
-            GenerateRequest,
-            GenerateResponse,
-            WrappedGenerateResponse,
-        )
-
-        seed = int(seed)
-        if seed != 0:
-            set_seed(seed)
-            logger.warning(f"set seed: {seed}")
-
-        # Parse reference audio aka prompt
-        prompt_tokens = encode_reference(
-            decoder_model=self._model,
-            reference_audio=reference_audio,
-            enable_reference_audio=enable_reference_audio,
-        )
-
-        # LLAMA Inference
-        request = dict(
-            device=self._model.device,
-            max_new_tokens=max_new_tokens,
-            text=text,
-            top_p=top_p,
-            repetition_penalty=repetition_penalty,
-            temperature=temperature,
-            compile=self._kwargs.get("compile", False),
-            iterative_prompt=chunk_length > 0,
-            chunk_length=chunk_length,
-            max_length=2048,
-            prompt_tokens=prompt_tokens if enable_reference_audio else None,
-            prompt_text=reference_text if enable_reference_audio else None,
-        )
-
-        response_queue = queue.Queue()
-        self._llama_queue.put(
-            GenerateRequest(
-                request=request,
-                response_queue=response_queue,
-            )
+        self._engine = TTSInferenceEngine(
+            self._llama_queue, self._model, precision, enable_compile
         )
 
-        segments = []
-
-        while True:
-            result: WrappedGenerateResponse = response_queue.get()
-            if result.status == "error":
-                raise result.response
-
-            result: GenerateResponse = result.response
-            if result.action == "next":
-                break
-
-            with autocast_exclude_mps(
-                device_type=self._model.device.type,
-                dtype=self._kwargs.get("precision", torch.bfloat16),
-            ):
-                fake_audios = decode_vq_tokens(
-                    decoder_model=self._model,
-                    codes=result.codes,
-                )
-
-            fake_audios = fake_audios.float().cpu().numpy()
-            segments.append(fake_audios)
-
-            if streaming:
-                yield fake_audios, None, None
-
-        if len(segments) == 0:
-            raise Exception("No audio generated, please check the input text.")
-
-        # No matter streaming or not, we need to return the final audio
-        audio = np.concatenate(segments, axis=0)
-        yield None, (self._model.spec_transform.sample_rate, audio), None
-
-        if torch.cuda.is_available():
-            torch.cuda.empty_cache()
-            gc.collect()
-
     def speech(
         self,
         input: str,
@@ -211,21 +124,31 @@ class FishSpeechModel:
         if speed != 1.0:
             logger.warning("Fish speech does not support setting speed: %s.", speed)
         import torchaudio
+        from tools.schema import ServeReferenceAudio, ServeTTSRequest
 
         prompt_speech = kwargs.get("prompt_speech")
-        result = self._inference(
-            text=input,
-            enable_reference_audio=kwargs.get(
-                "enable_reference_audio", prompt_speech is not None
-            ),
-            reference_audio=prompt_speech,
-            reference_text=kwargs.get("reference_text", ""),
-            max_new_tokens=kwargs.get("max_new_tokens", 1024),
-            chunk_length=kwargs.get("chunk_length", 200),
-            top_p=kwargs.get("top_p", 0.7),
-            repetition_penalty=kwargs.get("repetition_penalty", 1.2),
-            temperature=kwargs.get("temperature", 0.7),
-            streaming=stream,
+        prompt_text = kwargs.get("prompt_text", kwargs.get("reference_text", ""))
+        if prompt_speech is not None:
+            r = ServeReferenceAudio(audio=prompt_speech, text=prompt_text)
+            references = [r]
+        else:
+            references = []
+
+        assert self._engine is not None
+        result = self._engine.inference(
+            ServeTTSRequest(
+                text=input,
+                references=references,
+                reference_id=kwargs.get("reference_id"),
+                seed=kwargs.get("seed"),
+                max_new_tokens=kwargs.get("max_new_tokens", 1024),
+                chunk_length=kwargs.get("chunk_length", 200),
+                top_p=kwargs.get("top_p", 0.7),
+                repetition_penalty=kwargs.get("repetition_penalty", 1.2),
+                temperature=kwargs.get("temperature", 0.7),
+                streaming=stream,
+                format=response_format,
+            )
         )
 
         if stream:
@@ -241,7 +164,9 @@ class FishSpeechModel:
                     last_pos = 0
                     with writer.open():
                         for chunk in result:
-                            chunk = chunk[0]
+                            if chunk.code == "final":
+                                continue
+                            chunk = chunk.audio[1]
                             if chunk is not None:
                                 chunk = chunk.reshape((chunk.shape[0], 1))
                                 trans_chunk = torch.from_numpy(chunk)
@@ -256,7 +181,7 @@ class FishSpeechModel:
             return _stream_generator()
         else:
             result = list(result)
-            sample_rate, audio = result[0][1]
+            sample_rate, audio = result[0].audio
             audio = np.array([audio])
 
             # Save the generated audio

xinference/model/audio/melotts.py

@@ -0,0 +1,110 @@
+# Copyright 2022-2023 XProbe Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import logging
+from io import BytesIO
+from typing import TYPE_CHECKING, Optional
+
+from ...device_utils import get_available_device, is_device_available
+
+if TYPE_CHECKING:
+    from .core import AudioModelFamilyV1
+
+logger = logging.getLogger(__name__)
+
+
+class MeloTTSModel:
+    def __init__(
+        self,
+        model_uid: str,
+        model_path: str,
+        model_spec: "AudioModelFamilyV1",
+        device: Optional[str] = None,
+        **kwargs,
+    ):
+        self._model_uid = model_uid
+        self._model_path = model_path
+        self._model_spec = model_spec
+        self._device = device
+        self._model = None
+        self._kwargs = kwargs
+
+    @property
+    def model_ability(self):
+        return self._model_spec.model_ability
+
+    def load(self):
+        if self._device is None:
+            self._device = get_available_device()
+        else:
+            if not is_device_available(self._device):
+                raise ValueError(f"Device {self._device} is not available!")
+
+        import os
+        import sys
+
+        import nltk
+
+        # English language requires download averaged_perceptron_tagger_eng
+        nltk.download("averaged_perceptron_tagger_eng")
+
+        # The yaml config loaded from model has hard-coded the import paths. please refer to: load_hyperpyyaml
+        sys.path.insert(0, os.path.join(os.path.dirname(__file__), "../../thirdparty"))
+
+        from melo.api import TTS
+
+        config_path = os.path.join(self._model_path, "config.json")
+        ckpt_path = os.path.join(self._model_path, "checkpoint.pth")
+        self._model = TTS(
+            language=self._model_spec.language,
+            device=self._device,
+            config_path=config_path,
+            ckpt_path=ckpt_path,
+        )
+
+    def speech(
+        self,
+        input: str,
+        voice: str,
+        response_format: str = "mp3",
+        speed: float = 1.0,
+        stream: bool = False,
+        **kwargs,
+    ):
+        import soundfile
+
+        if stream:
+            raise Exception("MeloTTS does not support stream mode.")
+        assert self._model is not None
+        speaker_ids = self._model.hps.data.spk2id
+        if not voice:
+            voice = next(iter(speaker_ids.keys()))
+            logger.info("Auto select speaker: %s", voice)
+        elif voice not in speaker_ids:
+            raise ValueError(
+                f"Invalid voice: {voice}, available speakers: {speaker_ids}"
+            )
+        audio = self._model.tts_to_file(
+            text=input, speaker_id=speaker_ids[voice], speed=speed, **kwargs
+        )
+        # Save the generated audio
+        with BytesIO() as out:
+            with soundfile.SoundFile(
+                out,
+                "w",
+                self._model.hps.data.sampling_rate,
+                1,
+                format=response_format.upper(),
+            ) as f:
+                f.write(audio)
+            return out.getvalue()

xinference/model/audio/model_spec.json

@@ -236,11 +236,107 @@
     "multilingual": true
   },
   {
-    "model_name": "FishSpeech-1.4",
+    "model_name": "CosyVoice2-0.5B",
+    "model_family": "CosyVoice",
+    "model_id": "mrfakename/CosyVoice2-0.5B",
+    "model_revision": "5676baabc8a76dc93ef60a88bbd2420deaa2f644",
+    "model_ability": "text-to-audio",
+    "multilingual": true
+  },
+  {
+    "model_name": "FishSpeech-1.5",
     "model_family": "FishAudio",
-    "model_id": "fishaudio/fish-speech-1.4",
-    "model_revision": "069c573759936b35191d3380deb89183c0656f59",
+    "model_id": "fishaudio/fish-speech-1.5",
+    "model_revision": "268b6ec86243dd683bc78dab7e9a6cedf9191f2a",
+    "model_ability": "text-to-audio",
+    "multilingual": true
+  },
+  {
+    "model_name": "F5-TTS",
+    "model_family": "F5-TTS",
+    "model_id": "SWivid/F5-TTS",
+    "model_revision": "4dcc16f297f2ff98a17b3726b16f5de5a5e45672",
     "model_ability": "text-to-audio",
     "multilingual": true
+  },
+  {
+    "model_name": "F5-TTS-MLX",
+    "model_family": "F5-TTS-MLX",
+    "model_id": "lucasnewman/f5-tts-mlx",
+    "model_revision": "7642bb232e3fcacf92c51c786edebb8624da6b93",
+    "model_ability": "text-to-audio",
+    "multilingual": true
+  },
+  {
+    "model_name": "MeloTTS-English",
+    "model_family": "MeloTTS",
+    "model_id": "myshell-ai/MeloTTS-English",
+    "model_revision": "bb4fb7346d566d277ba8c8c7dbfdf6786139b8ef",
+    "model_ability": "text-to-audio",
+    "multilingual": false,
+    "language": "EN"
+  },
+  {
+    "model_name": "MeloTTS-English-v2",
+    "model_family": "MeloTTS",
+    "model_id": "myshell-ai/MeloTTS-English-v2",
+    "model_revision": "a53e3509c4ee4ff16d79272feb2474ff864e18f3",
+    "model_ability": "text-to-audio",
+    "multilingual": false,
+    "language": "EN"
+  },
+  {
+    "model_name": "MeloTTS-English-v3",
+    "model_family": "MeloTTS",
+    "model_id": "myshell-ai/MeloTTS-English-v3",
+    "model_revision": "f7c4a35392c0e9be24a755f1edb4c3f63040f759",
+    "model_ability": "text-to-audio",
+    "multilingual": false,
+    "language": "EN"
+  },
+  {
+    "model_name": "MeloTTS-French",
+    "model_family": "MeloTTS",
+    "model_id": "myshell-ai/MeloTTS-French",
+    "model_revision": "1e9bf590262392d8bffb679b0a3b0c16b0f9fdaf",
+    "model_ability": "text-to-audio",
+    "multilingual": false,
+    "language": "FR"
+  },
+  {
+    "model_name": "MeloTTS-Japanese",
+    "model_family": "MeloTTS",
+    "model_id": "myshell-ai/MeloTTS-Japanese",
+    "model_revision": "367f8795464b531b4e97c1515bddfc1243e60891",
+    "model_ability": "text-to-audio",
+    "multilingual": false,
+    "language": "JP"
+  },
+  {
+    "model_name": "MeloTTS-Spanish",
+    "model_family": "MeloTTS",
+    "model_id": "myshell-ai/MeloTTS-Spanish",
+    "model_revision": "dbb5496df39d11a66c1d5f5a9ca357c3c9fb95fb",
+    "model_ability": "text-to-audio",
+    "multilingual": false,
+    "language": "ES"
+  },
+  {
+    "model_name": "MeloTTS-Chinese",
+    "model_family": "MeloTTS",
+    "model_id": "myshell-ai/MeloTTS-Chinese",
+    "model_revision": "af5d207a364ea4208c6f589c89f57f88414bdd16",
+    "model_ability": "text-to-audio",
+    "multilingual": false,
+    "language": "ZH"
+  },
+  {
+    "model_name": "MeloTTS-Korean",
+    "model_family": "MeloTTS",
+    "model_id": "myshell-ai/MeloTTS-Korean",
+    "model_revision": "0207e5adfc90129a51b6b03d89be6d84360ed323",
+    "model_ability": "text-to-audio",
+    "multilingual": false,
+    "language": "KR"
   }
 ]

xinference/model/audio/model_spec_modelscope.json

@@ -17,6 +17,15 @@
     "model_ability": "audio-to-text",
     "multilingual": true
   },
+  {
+    "model_name": "Belle-whisper-large-v3-zh",
+    "model_family": "whisper",
+    "model_hub": "modelscope",
+    "model_id": "Xorbits/Belle-whisper-large-v3-zh",
+    "model_revision": "master",
+    "model_ability": "audio-to-text",
+    "multilingual": false
+  },
   {
     "model_name": "SenseVoiceSmall",
     "model_family": "funasr",
@@ -73,5 +82,23 @@
     "model_revision": "master",
     "model_ability": "text-to-audio",
     "multilingual": true
+  },
+  {
+    "model_name": "CosyVoice2-0.5B",
+    "model_family": "CosyVoice",
+    "model_hub": "modelscope",
+    "model_id": "iic/CosyVoice2-0.5B",
+    "model_revision": "master",
+    "model_ability": "text-to-audio",
+    "multilingual": true
+  },
+  {
+    "model_name": "F5-TTS",
+    "model_family": "F5-TTS",
+    "model_hub": "modelscope",
+    "model_id": "SWivid/F5-TTS_Emilia-ZH-EN",
+    "model_revision": "master",
+    "model_ability": "text-to-audio",
+    "multilingual": true
   }
 ]

xinference/model/audio/utils.py

@@ -11,8 +11,40 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+
+import io
+
+import numpy as np
+
 from .core import AudioModelFamilyV1
 
 
 def get_model_version(audio_model: AudioModelFamilyV1) -> str:
     return audio_model.model_name
+
+
+def ensure_sample_rate(
+    audio: np.ndarray, old_sample_rate: int, sample_rate: int
+) -> np.ndarray:
+    import soundfile as sf
+    from scipy.signal import resample
+
+    if old_sample_rate != sample_rate:
+        # Calculate the new data length
+        new_length = int(len(audio) * sample_rate / old_sample_rate)
+
+        # Resample the data
+        resampled_data = resample(audio, new_length)
+
+        # Use BytesIO to save the resampled data to memory
+        with io.BytesIO() as buffer:
+            # Write the resampled data to the memory buffer
+            sf.write(buffer, resampled_data, sample_rate, format="WAV")
+
+            # Reset the buffer position to the beginning
+            buffer.seek(0)
+
+            # Read the data from the memory buffer
+            audio, sr = sf.read(buffer, dtype="float32")
+
+    return audio

xinference/model/audio/whisper.py

@@ -13,9 +13,12 @@
 # limitations under the License.
 import logging
 import os
+import typing
 from glob import glob
 from typing import TYPE_CHECKING, Dict, List, Optional, Union
 
+from typing_extensions import TypedDict
+
 from ...device_utils import (
     get_available_device,
     get_device_preferred_dtype,
@@ -28,6 +31,13 @@ if TYPE_CHECKING:
 logger = logging.getLogger(__name__)
 
 
+class WhisperModelConfig(TypedDict, total=False):
+    chunk_length_s: Optional[float]
+    stride_length_s: Optional[float]
+    return_timestamps: Optional[bool]
+    batch_size: Optional[int]
+
+
 class WhisperModel:
     def __init__(
         self,
@@ -35,6 +45,7 @@ class WhisperModel:
         model_path: str,
         model_spec: "AudioModelFamilyV1",
         device: Optional[str] = None,
+        max_new_tokens: Optional[int] = 128,
         **kwargs,
     ):
         self._model_uid = model_uid
@@ -42,7 +53,21 @@ class WhisperModel:
         self._model_spec = model_spec
         self._device = device
         self._model = None
-        self._kwargs = kwargs
+        self._max_new_tokens = max_new_tokens
+        self._model_config: WhisperModelConfig = self._sanitize_model_config(
+            typing.cast(WhisperModelConfig, kwargs)
+        )
+
+    def _sanitize_model_config(
+        self, model_config: Optional[WhisperModelConfig]
+    ) -> WhisperModelConfig:
+        if model_config is None:
+            model_config = WhisperModelConfig()
+        model_config.setdefault("chunk_length_s", 30)
+        model_config.setdefault("stride_length_s", None)
+        model_config.setdefault("return_timestamps", False)
+        model_config.setdefault("batch_size", 16)
+        return model_config
 
     @property
     def model_ability(self):
@@ -75,10 +100,10 @@ class WhisperModel:
             model=model,
             tokenizer=processor.tokenizer,
             feature_extractor=processor.feature_extractor,
-            max_new_tokens=128,
-            chunk_length_s=30,
-            batch_size=16,
-            return_timestamps=False,
+            chunk_length_s=self._model_config.get("chunk_length_s"),
+            stride_length_s=self._model_config.get("stride_length_s"),
+            return_timestamps=self._model_config.get("return_timestamps"),
+            batch_size=self._model_config.get("batch_size"),
             torch_dtype=torch_dtype,
             device=self._device,
         )
@@ -185,13 +210,13 @@ class WhisperModel:
             logger.warning(
                 "Prompt for whisper transcriptions will be ignored: %s", prompt
             )
+        generate_kwargs = {"max_new_tokens": self._max_new_tokens, "task": "transcribe"}
+        if language is not None:
+            generate_kwargs["language"] = language
+
         return self._call_model(
             audio=audio,
-            generate_kwargs=(
-                {"language": language, "task": "transcribe"}
-                if language is not None
-                else {"task": "transcribe"}
-            ),
+            generate_kwargs=generate_kwargs,
             response_format=response_format,
             temperature=temperature,
             timestamp_granularities=timestamp_granularities,